README.md

---
license: apache-2.0
language:
- en
pretty_name: Reasoning and Tool Calling
task_categories:
- text-generation
tags:
- reasoning
- tool-calling
- function-calling
configs:
- config_name: conversations
  default: true
  data_files:
  - split: train
    path: data/conversations/train-*.parquet
  - split: test
    path: data/conversations/test-*.parquet
- config_name: calibration_histories
  data_files:
  - split: calibration
    path: data/calibration_histories/calibration.parquet
- config_name: calibration_windows
  data_files:
  - split: calibration
    path: data/calibration_windows/calibration.parquet
---

# Reasoning and Tool Calling

This dataset converts tagged reasoning and tool-use conversations into typed
messages and tool definitions. It also includes calibration data derived from
those converted conversations.

The source is
[`Mustafaege/qwen3.5-toolcalling-v2`](https://huggingface.co/datasets/Mustafaege/qwen3.5-toolcalling-v2)
at revision `8f0343a5613879fefda0eb002d10ff7150a2c588`.

## How this differs from the source

The source stores 92,153 train conversations and 10,240 test conversations in
a tagged message format. Protocol instructions, tool declarations, reasoning,
calls, and tool results can appear inside message text.

This release keeps rows that can be converted to one typed tool-calling schema
without guessing about call order, tool identity, or message boundaries. It
contains 56,032 train rows and 6,243 test rows.

| Property | Source | This release |
|---|---|---|
| Train rows | 92,153 | 56,032 |
| Test rows | 10,240 | 6,243 |
| Message form | Tagged text | Typed messages with `reasoning_content` and `tool_calls` |
| Tool declarations | Embedded in protocol text | Structured `tools` field |
| Tool arguments | Embedded JSON | Canonical JSON string |
| Tool results | Tagged text | `tool` messages linked by call identifier |
| Row identity | Source position | Source revision, file hash, row number, and content hash |
| Calibration data | Not included | 25 rendered histories and four token windows |

The conversion does not add conversations. It extracts protocol structure,
normalizes the representation, and omits rows whose structure cannot be mapped
under the rules below.

A row enters `conversations` when all of these conditions hold:

- it contains at least one tool call;
- tool declarations can be extracted without ambiguity;
- every called tool has a matching declaration;
- call arguments contain valid JSON;
- tool results can be paired with calls in sequence;
- reasoning and answer text occupy valid positions in that sequence;
- source-format control tags do not remain after conversion.

The main primary exclusion reasons were:

| Primary reason | Train | Test |
|---|---:|---:|
| Assistant content appears before pending tool results | 15,120 | 1,667 |
| No tool call | 10,745 | 1,176 |
| An unresolved call precedes another call | 4,650 | 534 |
| Embedded tool declarations are ambiguous | 2,194 | 251 |
| Empty tool response | 1,378 | 163 |
| Other tag, JSON, tool-name, or sequence failures | 2,034 | 206 |
| **Rows omitted** | **36,121** | **3,997** |

## Configurations

| Configuration | Split | Rows | Intended use |
|---|---|---:|---|
| `conversations` | `train` | 56,032 | Training, calibration-candidate selection, or format study |
| `conversations` | `test` | 6,243 | The same conversion applied to the source test split |
| `calibration_histories` | `calibration` | 25 | Inspecting the selected rendered histories and their token IDs |
| `calibration_windows` | `calibration` | 4 | Reusing the exact four 2,048-token calibration windows |

The train and test names preserve the source assignments.

## Conversation representation

Each row in `conversations` contains:

- source dataset, revision, split, file, file SHA-256, row number, and content
  identity SHA-256;
- style and tool-call category;
- typed messages;
- typed tool declarations;
- counts for source messages, users, reasoning messages, answers, tools, calls,
  results, and terminal unresolved calls.

Assistant messages can contain:

- `content` for visible answer text;
- `reasoning_content` for extracted reasoning;
- `tool_calls` with call ID, type, tool name, and `arguments_json`.

Tool messages carry `tool_call_id`. Tool declarations contain name,
description, and `parameters_json`.

`arguments_json` and `parameters_json` remain strings because argument values
and JSON Schema properties have heterogeneous types across tools. Decode them
with a JSON parser when object values are needed.

## Calibration views

The `conversations` configuration does not depend on a model tokenizer. The two
calibration configurations record a DeepSeek-V4-Flash-0731 rendering and use
its tokenizer. They are included for users who need the exact rendered input;
other uses should start from `conversations`.

The calibration view was formed from 25 accepted train histories:

| Category | Histories |
|---|---:|
| One tool call | 9 |
| Multiple tool calls | 8 |
| Reasoning with at most three tool calls | 5 |
| Reasoning with more than three tool calls | 3 |

The histories were rendered with thinking enabled and DSML tool calls. They
were tokenized and concatenated within each category, and one 2,048-token slice
was taken from each category stream. `calibration_windows` contains those four
slices, for 8,192 tokens in total.

`calibration_histories` stores each selected conversation, its rendered prompt,
token IDs, category, stream offsets, and content hashes.

`calibration_windows` stores each slice's token IDs, category, stream offsets,
marker counts, selection rule, and contributing history identities. Tokenizer
hashes identify the vocabulary used for these token IDs.

## License and modifications

The source declares the Apache License 2.0. This release uses the same license.